from rank_bm25 import BM25Plus import datasets from sklearn.base import BaseEstimator from sklearn.model_selection import GridSearchCV from huggingface_hub import create_repo from huggingface_hub.utils._errors import HfHubHTTPError N_NEGATIVE_DOCS = 10 # Prepare documents def create_text(example:dict) -> str: return "\n".join([example["section"], example["title"], example["content"]]) documents = datasets.load_dataset("lyon-nlp/mteb-fr-retrieval-syntec-s2p", "documents")["test"] documents = documents.add_column("text", [create_text(x) for x in documents]) documents = documents.rename_column("id", "doc_id") documents = documents.remove_columns(["url", "title", "section", "content"]) # Prepare queries queries = datasets.load_dataset("lyon-nlp/mteb-fr-retrieval-syntec-s2p", "queries")["test"] queries = queries.rename_columns({"Question": "queries", "Article": "doc_id"}) queries = queries.map(lambda x: {"doc_id": [x["doc_id"]]}) # Optimize BM25 parameters ### Build sklearn estimator feature BM25 class BM25Estimator(BaseEstimator): def __init__(self, corpus_dataset:datasets.Dataset, *, k1:float=1.5, b:float=.75, delta:int=1): """Initialize BM25 estimator using the coprus dataset. The dataset must contain 2 columns: - "doc_id" : the documents ids - "text" : the document texts Args: corpus_dataset (datasets.Dataset): _description_ k1 (float, optional): _description_. Defaults to 1.5. b (float, optional): _description_. Defaults to .75. delta (int, optional): _description_. Defaults to 1. """ self.is_fitted_ = False self.corpus_dataset = corpus_dataset self.k1 = k1 self.b = b self.delta=delta self.bm25 = None def tokenize_corpus(self, corpus:list[str]) -> list[str]: """Tokenize a corpus of strings Args: corpus (list[str]): the list of string to tokenize Returns: list[str]: the tokeinzed corpus """ if isinstance(corpus, str): return corpus.lower().split() return [c.lower().split() for c in corpus] def fit(self, X=None, y=None): """Fits the BM25 using the dataset of documents Args are placeholders required by sklearn """ tokenized_corpus = self.tokenize_corpus(self.corpus_dataset["text"]) self.bm25 = BM25Plus( corpus=tokenized_corpus, k1=self.k1, b=self.b, delta=self.delta ) self.is_fitted_ = True return self def predict(self, query:str, topN:int=10) -> list[str]: """Returns the best doc ids in order of best relevance first Args: query (str): _description_ topN (int, optional): _description_. Defaults to 10. Returns: list[str]: _description_ """ if not self.is_fitted_: self.fit() tokenized_query = self.tokenize_corpus(query) best_docs = self.bm25.get_top_n(tokenized_query, self.corpus_dataset["text"], n=topN) best_docs_ids = [self.corpus_dataset["doc_id"][self.corpus_dataset["text"].index(doc)] for doc in best_docs] return best_docs_ids def score(self, queries:list[str], relevant_docs:list[list[str]]): """Scores the bm25 using the queries and relevant docs, using MRR as the metric. Args: queries (list[str]): list of queries relevant_docs (list[list[str]]): list of relevant documents ids for each query """ best_docs_ids_preds = [self.predict(q, len(self.corpus_dataset)) for q in queries] best_docs_isrelevant = [ [ doc in rel_docs for doc in best_docs_ids_pred ] for best_docs_ids_pred, rel_docs in zip(best_docs_ids_preds, relevant_docs) ] mrrs = [self._compute_mrr(preds) for preds in best_docs_isrelevant] mrr = sum(mrrs)/len(mrrs) return mrr def _compute_mrr(self, predictions:list[bool]) -> float: """Compute the mrr considering a list of boolean predictions. Example: if predictions = [False, False, True, False], it would indicate that only the third document was labeled as relevant to the query Args: predictions (list[bool]): the binarized relevancy of predictions Returns: float: the mrr """ if any(predictions): mrr = [1/(i+1) for i, pred in enumerate(predictions) if pred] mrr = sum(mrr)/len(mrr) return mrr else: return 0 ### Perform gridSearch to find best parameters for BM25 print("Optimizing BM25 parameters...") params = { "k1":[1., 1.25, 1.5, 1.75], "b": [.5, .75, 1.], "delta": [0, 1, 2] } gscv = GridSearchCV(BM25Estimator(documents), params) gscv.fit(queries["queries"], queries["doc_id"]) print("Best parameterss :", gscv.best_params_) print("Best MRR score :", gscv.best_score_) # Build reranking dataset with positives and negative queries using best estimator reranking_dataset = datasets.Dataset.from_dict( { "query": queries["queries"], "positive": queries["doc_id"], "negative": [ [doc_id for doc_id in gscv.estimator.predict(q, N_NEGATIVE_DOCS) if doc_id not in relevant_ids] for q, relevant_ids in zip(queries["queries"], queries["doc_id"]) ] }) # Push dataset to hub ### create HF repo repo_id = "lyon-nlp/mteb-fr-reranking-syntec-s2p" try: create_repo(repo_id, repo_type="dataset") except HfHubHTTPError as e: print("HF repo already exist") ### push to hub reranking_dataset.push_to_hub(repo_id, config_name="queries", split="test") documents.push_to_hub(repo_id, config_name="documents", split="test")